Police know that most crimes are committed by repeat offenders, so a staple of police work is identifying patterns that link crimes together and deduce the results to specific individuals. This becomes difficult when analysts look at thousands of cases each year.
Though the human brain excels at identifying patterns, it can succumb to information overload.
Chicago Police Department crime analysts sometimes look at 100 cases a week, sifting through hundreds of data elements trying to unravel patterns that might lead to a break. It requires a lot of legwork and some luck. However, police brass hope a new neural network system will take some chance out of the equation, and add more of a scientific formula to solving crimes.
"The average cases kind of blend together after a while," said Steve Maris, assistant director for Information Services for the Chicago Police Department. "This would be able to segment those for us."
The Classification System for Serial Criminal Patterns (CSSCP) is the brainchild of Dr. Tom Muscarello, an assistant professor at DePaul University.
It's different from other crime-analysis systems being used by law enforcement in that the CSSCP thinks 24 hours a day, seven days a week -- not just when prompted by an analyst. It can, however, be prompted to search for a particular data set, analyze data from multiple crimes and find patterns that link crimes without human intervention.
Running 24/7, the system combs through police department IT systems, searching for patterns or clusters of data elements that might tie together a string of crimes and give police the data they need to find the perpetrators. The system assigns numerical values to different data elements in each crime, including crime type, suspect description and profile, getaway vehicle and so forth.
The system uses pattern-recognition software that is "trained" to find those clusters of data.
Neural networks are considered artificial intelligence -- the networks attempt to imitate the human brain in the way the brain programs data structures and recognizes patterns. Neural networks function by creating connections between processing elements, which are the equivalent of neurons to the computer system.
These networks become adept at predicting events when they have a large database of examples from which to draw, and are typically "trained" by being fed large amounts of data and "taught" rules about interpreting relationships between that data.
"It cuts down on manual intervention," Maris said. "[A detective] reads 100 cases this week, he reads 100 next week. Can he go back and remember which case belongs where?
"Right now, we have a lack of computer tools to assist the crime analyst," he continued. "Crime analysts are doing a lot of legwork -- reading lots of cases using text searches to find cases. Nothing is grouping the case by offender patterns, MO [Modus Operandi], things like that."
Also, since cases are often assigned arbitrarily, communication between detectives may not be what it should, and links between cases may go uncovered without such a system.
"Some sergeant is passing [case assignments] out to people," Muscarello said. "If [the sergeant] doesn't know right off the top of his head that it sounds like a case that's related to others, it's kind of a round-robin thing; 'Well, Joe you got two yesterday so I'm going to give these to Frank.' A lot of times people don't communicate with each other as often as you'd think people in an office environment would."
Modeling the Brain
The CSSCP has been in the works for a decade, and after considerable tinkering, should be ready for the Chicago Police Department this year. Muscarello said changes in leadership at the police department and adjustments to the system have delayed its advance.
The neural network was derived from analysis of the most successful detectives in Chicago. These six detectives were at the top of the department in terms of arrests made and cases closed, Muscarello said.
"We picked their brains for the type of patterns they were looking for. We looked at what they did, and found there was no one way they did their work," he said. "Some of them concentrated on the victim, some on the time of day, but they all concentrated on something, and it helped them solve the crime. We picked out the best data features to look at and tried to normalize them."
What he means by "normalize" is programming the system to look for patterns the way the human brain does. Take height, a common data element, for instance.
Eyewitness accounts are notoriously inaccurate, so trying to be too detailed can lead a detective in the wrong direction. Muscarello said in terms of height, people think in terms of tall, average and short -- and that's how it will be programmed into the network.
"The victim has very little time to see the offender," Muscarello said. "Even in the best of circumstances, people are usually off when they try to estimate someone's height unless they're about your height."
One of the six detectives focused heavily on getaway vehicles. Like the height of suspects, Muscarello said, this data is "normalized" when programmed into the system. A victim might describe a getaway vehicle as a navy blue Toyota Corolla, but focusing just on that type of vehicle might lead an investigator to a dead end.
"That's too exact," Muscarello said.
A good investigator would focus on a dark vehicle, probably foreign, maybe Japanese. That's how the system is programmed to recognize a cluster or a pattern. The detective can find such a cluster by clicking on a drop-down box on the computer or typing in a query.
"The way we built this is the network will know the important things to look at, and it would also learn the less important things," Muscarello said. "So on its own it would do a pass with what it learned was important. What we also did is give the computer the capability so that the interface allowed you to either access all of the pre-determined case clusters [that the system is programmed to recognize], or enter your new data select things you were interested in looking for."
The department will test linking the CSSCP to the Citizen Law Enforcement Analysis and Reporting (CLEAR) system, the state's crime data warehouse.
Both resources could help police solve crimes in two direct ways. First, by locating clusters of data elements that illustrate a clear pattern and point to a specific suspect(s); and second, by having the CSSCP link the detainee to previous crimes or even cold cases.
With a suspect in custody, police can examine how the crime was conducted, then sift through the CSSCP and try to match the characteristics of the latest crime with ones from the past, Maris said.
When they find a pattern, they can interrogate the suspect further with the evidence.
"'We caught you for this burglary; did you do these other six burglaries, too? You used the same MO, the same characteristics,'" Maris explained as an example of tying a known suspect to previous cases.
Most convicted criminals that are incarcerated continue committing crimes once released, so going back and looking at cold cases or even solved ones can lead to new arrests.
"We know that most crimes are committed by a few criminals, and we just aren't closing out that many cases," Muscarello said.
He's spent the last decade trying to fix that.